Species distribution models predict range expansion better than chance but not better than a simple dispersal model

Abstract

The evaluation of species distribution models (SDMs) is a crucial step; usually, a random subsample of data is used to test prediction capacity. This procedure, called cross-validation, has been recently shown to overestimate SDMs performance due to spatial autocorrelation. In the case of expanding species, there exists the possibility to test the predictions with non-random geographically structured data, i.e., a new data set which corresponds to the last occupied localities. The aim of this study was to evaluate the capacity of SDMs to predict the range expansion pattern of six free-living deer species in Great Britain and to assess whether SDMs perform better than a simple dispersal model – a null model that assumes no environmental control in the expansion process. Distribution data for the species prior to 1972 were used to train the SDMs (ENFA, MAXENT, logistic regression and an ensemble model) in order to obtain suitability maps. Additionally, the geographical distance to the localities occupied in 1972 was considered a proxy of the probability that a certain locality has to be occupied during an expansion process considering only dispersal (GD model). Subsequently, we analysed whether the species increased their ranges between 1972 and 2006 according to the estimated suitability patterns and whether or not SDMs predictions outperformed GD predictions. SDMs showed a high discrimination capacity in the training data, with the ensemble models performing the best and ENFA models the worst. SDMs predictions also worked better than chance in classifying new occupied localities, although differences among techniques disappeared and the predictions showed no difference with respect to GD. Spatial autocorrelation of both the environmental predictors and the expansion process may explain these results which illustrate that GD is a much more parsimonious model than any of the SDMs and may thus be preferable both for prediction and explanation. Overestimation of SDMs performance and usefulness may be a common fact.